Scores based on a WDL model vs the pawn value

Since the "Normalize evaluation" patch of Nov 5, 2022, the scores are based on a WDL model (based on fishtest's games = Stockfish selfplay).

A score of +1.00 means 50% chance of winning, -1.00 means 50% chance of losing.

The WDL model of Stockfish is not very accurate :


After a "go depth 30 searchmoves g2g4" command, Stockfish 15.1 evaluates 1.g4 at -1.19 so 60% chance of losing.
But after playing 88k real games, the WDL ratio is 95% chance of losing.